Intelligent Manufacturing Operation Management Methods and Systems Based on Industrial Internet of Things

Abstract

This invention discloses a method and system for intelligent manufacturing operation management based on the Industrial Internet of Things (IIoT), belonging to the field of IIoT. The method includes: collecting multi-dimensional manufacturing operation data through IIoT sensing; performing edge processing on the multi-dimensional manufacturing operation data to obtain a standardized manufacturing data stream; mapping the standardized manufacturing data stream to a manufacturing semantic knowledge graph to construct a multi-level manufacturing operation data model; performing collaborative optimization of manufacturing operations based on the multi-level manufacturing operation data model to formulate an intelligent decision instruction set; executing the intelligent decision instruction set, monitoring execution feedback data in real time, and performing closed-loop updates based on the execution feedback data. This application solves the technical problems of low production operation efficiency and insufficient system intelligence in existing intelligent manufacturing operation management, achieving efficient, collaborative, and intelligent control of intelligent manufacturing operation management, and reducing operating costs.

Description

Technical Field

[0001] This invention relates to the field of industrial Internet of Things (IIoT), and more specifically to intelligent manufacturing operation management methods and systems based on IIoT. Background Technology

[0002] Intelligent manufacturing is the core direction for the digital, networked, and intelligent transformation and upgrading of the manufacturing industry. Industrial Internet of Things (IIoT) technology, as the core support for intelligent manufacturing, has been applied on a large scale in mainstream industrial scenarios such as discrete manufacturing and process manufacturing. Currently, existing intelligent manufacturing operation and management solutions still have many shortcomings, making it difficult to meet the operational needs of modern manufacturing for efficient collaboration, precise control, and intelligent decision-making: the quality of data collected from industrial sites cannot meet the requirements of upper-level intelligent modeling and decision analysis; massive amounts of on-site operational data are disconnected from production business entities and business logic, forming numerous data silos, making it impossible to achieve self-optimization of decisions based on actual execution results. Summary of the Invention

[0003] This application provides a method and system for intelligent manufacturing operation management based on the Industrial Internet of Things, aiming to solve the technical problems of low production and operation efficiency and insufficient system intelligence in existing intelligent manufacturing operation management.

[0004] In view of the above problems, this application provides a method and system for intelligent manufacturing operation management based on the Industrial Internet of Things.

[0005] The first aspect disclosed in this application provides a smart manufacturing operation management method based on the Industrial Internet of Things, the method comprising: Multi-dimensional manufacturing operation data is collected through industrial IoT sensing, and edge processing is performed on the multi-dimensional manufacturing operation data to obtain a standardized manufacturing data stream. The standardized manufacturing data stream is mapped to a manufacturing semantic knowledge graph to construct a multi-level manufacturing operation data model. Based on the multi-level manufacturing operation data model, manufacturing operation collaborative optimization is performed to formulate an intelligent decision instruction set. The intelligent decision instruction set is executed, and the execution feedback data is monitored in real time. Closed-loop updates are performed based on the execution feedback data.

[0006] Another aspect of this application discloses a smart manufacturing operations management system based on the Industrial Internet of Things (IIoT), which includes: The processing module is used to collect multi-dimensional manufacturing operation data through industrial IoT sensing, perform edge processing on the multi-dimensional manufacturing operation data, and obtain a standardized manufacturing data stream; the mapping module is used to map the standardized manufacturing data stream to a manufacturing semantic knowledge graph to construct a multi-level manufacturing operation data model; the optimization module is used to perform collaborative optimization of manufacturing operations based on the multi-level manufacturing operation data model and formulate an intelligent decision instruction set; the execution module is used to execute the intelligent decision instruction set, monitor the execution feedback data in real time, and perform closed-loop updates based on the execution feedback data.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: Through a comprehensive technical solution encompassing edge data standardization processing, manufacturing semantic knowledge graph construction, multi-level operational data model mapping, multi-branch parallel collaborative optimization, and closed-loop execution updates, high-quality standardized transformation of multi-source manufacturing operation data has been achieved. This solution breaks down the semantic barriers between data and production business entities, enabling parallel collaboration and conflict-free decision output in task scheduling, resource allocation, and anomaly handling. It also establishes a quantitative monitoring system for decision execution deviations and a closed-loop self-optimization mechanism across the entire chain. This effectively improves the accuracy of production decisions and the speed of on-site execution response, increases the overall equipment utilization rate and the efficiency of production resource utilization, and achieves efficient, collaborative, and intelligent control of intelligent manufacturing operations management, thereby reducing operating costs.

[0008] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0009] Figure 1 A flowchart illustrating a smart manufacturing operation management method based on the Industrial Internet of Things is provided for embodiments of this application; Figure 2 This application provides a schematic diagram of the structure of an intelligent manufacturing operation management system based on the Industrial Internet of Things.

[0010] Explanation of reference numerals in the attached diagram: Processing module 11, Mapping module 12, Optimization module 13, Execution module 14. Detailed Implementation

[0011] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0012] The overall concept of the technical solution provided in this application is as follows: This application provides a method and system for intelligent manufacturing operation management based on the Industrial Internet of Things (IIoT). It achieves full-domain collection of multi-dimensional manufacturing operation data through an IIoT sensing system, and relies on edge gateway nodes to complete outlier detection, cleaning, completion, normalization, and structured encapsulation of the data, generating a standardized manufacturing data stream. Then, it uses a self-constructed manufacturing semantic knowledge graph to perform semantic association calculations between time-series data and business entities, building a multi-layered manufacturing operation data model that integrates static manufacturing knowledge and dynamic operation data. Based on this model, it achieves parallel collaborative optimization of three branches: task scheduling, resource allocation, and anomaly handling, generating a standardized intelligent decision instruction set. Simultaneously, through real-time collection and deviation quantification analysis of instruction execution feedback data, it triggers incremental learning signals when deviations exceed limits, completing a closed-loop update of the entire link model and algorithm.

[0013] After introducing the basic principles of this application, various non-limiting embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0014] Example 1, as Figure 1 As shown in the embodiments of this application, a smart manufacturing operation management method based on the Industrial Internet of Things is provided, the method comprising: S100: Collect multi-dimensional manufacturing operation data through industrial IoT sensing, perform edge processing on the multi-dimensional manufacturing operation data, and obtain a standardized manufacturing data stream.

[0015] Specifically, firstly, industrial IoT sensing terminals, including industrial sensors, programmable logic controllers (PLCs), RFID readers, smart meters, and visual acquisition terminals, are deployed at the smart manufacturing production site to collect multi-dimensional manufacturing operation data in real time. This multi-dimensional manufacturing operation data covers the entire smart manufacturing production chain, including multi-dimensional, multi-source, and heterogeneous operational data. It includes equipment status data characterizing the operating conditions of production equipment, process parameter data guiding production execution, material flow data recording the flow of production elements, and environmental monitoring data monitoring the conditions at the production site. The collection process relies on various industrial sensors, PLCs, RFID readers, smart meters, and visual acquisition terminals deployed at the industrial IoT sensing layer to achieve high-frequency, synchronous collection of operational data for the entire production process, from equipment operation and process execution to material flow and the site environment. After collection, the raw multi-dimensional manufacturing operation data is transmitted to the nearest industrial IoT edge gateway node to complete the full... Edge processing of the process involves edge gateway nodes, which are edge computing hardware nodes deployed between the perception layer and the upper platform layer at the production site. They possess capabilities for local data processing, low-latency transmission, and multi-industry protocol conversion. Edge processing is executed according to a standardized process to obtain aligned time-series data. Finally, the aligned time-series data is structured and encapsulated according to data type to generate a standardized manufacturing data stream. This standardized manufacturing data stream is a time-series data stream that has undergone the above-mentioned full-process edge processing, and has been encapsulated in accordance with the general technical specifications of industrial IoT. It has a unified data format, a unified timestamp identifier, a unified semantic encoding, and can be directly read by the upper-layer system and stably transmitted in a streaming manner. It adopts the JSON / OPCUA standardized data format commonly used in industrial scenarios. Each data entry contains eight mandatory fields: a globally unique data ID, a unique device identifier code, a data type label, a collection timestamp, a normalized value, a raw value, a unit of measurement, and a data quality label. It is compatible with mainstream industrial communication protocols.

[0016] S200: Map the standardized manufacturing data stream to a manufacturing semantic knowledge graph to construct a multi-level manufacturing operation data model.

[0017] Specifically, two types of feature extraction operations are performed simultaneously on the input standardized manufacturing data stream: first, equipment identification analysis, extracting multiple equipment identification tag data; and second, time feature analysis, extracting time series feature vectors. After the feature extraction is completed simultaneously, the construction of a manufacturing semantic knowledge graph is carried out. The manufacturing semantic knowledge graph is a semantic knowledge network used to structurally represent various entities and the relationships between entities in the entire intelligent manufacturing process, with the triple of head entity-relationship type-tail entity as the basic building block. After completing the construction of the manufacturing semantic knowledge graph and the extraction of two types of features, a semantic association mapping is performed based on the semantic feature vectors and the manufacturing semantic knowledge graph to construct a node-data association mapping matrix. Specifically, this involves: performing full node matching based on the semantic feature vectors and the manufacturing semantic knowledge graph to extract the embedding vectors of all nodes in the knowledge graph; calculating cosine similarity between the embedding vector and the semantic feature vector of each node to obtain multiple cosine similarity values. The cosine similarity calculation characterizes the semantic similarity between vectors by calculating the cosine of the angle between two vectors. Subsequently, a similarity threshold is preset, and the multiple cosine similarity values ​​are compared with the similarity threshold one by one. The process involves extracting target nodes with cosine similarity values ​​greater than a similarity threshold, identifying these target nodes as the first nodes semantically related to the time-series feature vectors, and iterating through the process until all nodes in the manufacturing semantic knowledge graph have been traversed and verified. This process yields multiple associated nodes that are semantically related to the time-series data. Finally, based on the multiple associated nodes and their corresponding cosine similarity values, a node-data association mapping matrix is ​​constructed, where rows correspond to knowledge graph entity nodes, columns correspond to time-series data features, and matrix values ​​correspond to semantic association degrees. This completes the quantitative association mapping between the standardized manufacturing data flow and the manufacturing semantic knowledge graph.

[0018] After completing the mapping matrix construction, the standardized manufacturing data flow is precisely embedded into the corresponding entity nodes of the manufacturing semantic knowledge graph according to the node-data association mapping matrix and the multiple equipment identification tags extracted in the early stage. This constructs a multi-level manufacturing operation data model. The multi-level manufacturing operation data model is a multi-level data model with global semantic association capabilities formed by deeply binding manufacturing operation data at different levels such as equipment layer, process layer, material layer, production line layer, and operation layer with the corresponding knowledge graph nodes based on the entity hierarchy structure of the manufacturing semantic knowledge graph.

[0019] S300: Based on the multi-level manufacturing operation data model, perform collaborative optimization of manufacturing operations and formulate intelligent decision instruction sets.

[0020] Specifically, firstly, based on the multi-level manufacturing operation data model, a comprehensive manufacturing operation collaborative optimization analysis is conducted. Combined with the business objectives of intelligent manufacturing operations, including on-time order delivery, maximizing resource utilization efficiency, rapid closure of production anomalies, and minimizing operating costs, the entire process of collaborative optimization is streamlined. The objectives, constraints, data input sources, and business linkage rules of each optimization stage are clarified, constructing standardized manufacturing operation collaborative optimization process data. This data represents a digital decomposition and optimization path definition of the entire intelligent manufacturing operation business process, used to standardize the execution boundaries and collaborative logic of subsequent optimization stages. Subsequently, this manufacturing operation collaborative optimization process data is decomposed into three fully parallel optimization branches: task scheduling optimization, resource allocation optimization, and anomaly handling optimization. These three branches synchronously retrieve data from the same source based on the same multi-level manufacturing operation data model to perform parallel optimization calculations, generating directly applicable... The system issues task scheduling decision instructions, resource allocation decision instructions, and exception handling decision instructions. After the three parallel optimization branches have completed the generation of their respective decision instructions, the task scheduling decision instructions, resource allocation decision instructions, and exception handling decision instructions are collaboratively verified and merged and sorted. The collaborative verification involves adjusting instructions with execution conflicts, resource contention, and timing contradictions to ensure that all instructions have consistent execution goals, conflict-free resource allocation, and self-consistent timing logic. The merging and sorting prioritizes instructions according to their urgency level, execution timing dependency, and scope of impact. High-level exception handling instructions take precedence over regular task scheduling and resource allocation instructions, and task scheduling instructions for preceding processes take precedence over instructions for subsequent processes. Ultimately, a unified, collaborative, conflict-free, and closed-loop executable intelligent decision instruction set is formed. This intelligent decision instruction set is a standardized set of executable instructions covering all scenarios of production task scheduling, dynamic resource allocation, and production exception handling.

[0021] S400: Execute the intelligent decision instruction set, monitor the execution feedback data in real time, and perform closed-loop updates based on the execution feedback data.

[0022] Specifically, the pre-generated intelligent decision-making instruction set is first distributed to the equipment controllers and material handling systems on the production floor via an industrial IoT edge gateway. This intelligent decision-making instruction set is a standardized set of instructions covering three business areas: task scheduling, resource allocation, and anomaly handling, which can be directly parsed by the underlying execution units. Equipment controllers refer to low-level control units such as PLCs and CNCs that can drive equipment actions. The material handling system refers to the in-plant logistics execution system composed of AGVs, automated warehouses, and roller conveyors. After the instructions are distributed, corresponding equipment control instructions and material handling instructions are triggered simultaneously, entering the execution phase. Throughout the instruction execution process, real-time data collection is continuously conducted based on the industrial IoT sensing system to obtain execution feedback data directly related to instruction execution. This execution feedback data is quantitative data characterizing the actual implementation effect of the decision, including three categories: actual equipment status change data, actual material flow data, and actual production cycle data, used to accurately reflect the differences between equipment, materials, production cycle, and expected targets. Subsequently, based on intelligent... The decision instruction set conducts execution expectation analysis, setting expected execution goals for each instruction, i.e., the ideal state and numerical benchmark that the instruction should achieve. Then, the execution feedback data is compared with the expected execution goals item by item, calculating multiple deviation components. Each deviation component represents the degree of execution deviation in a single dimension. The multiple deviation components are synthesized according to the data direction and business dimension to construct a deviation vector. This vector can uniformly represent the overall situation and source direction of multi-dimensional execution deviations. The length of the deviation vector is then calculated to obtain a comprehensive length value, which is normalized to generate a comprehensive deviation metric value. This value is used to quantitatively determine whether the overall deviation exceeds the standard. At the same time, a deviation threshold is preset as the upper limit of acceptable execution deviation. When the comprehensive deviation metric value exceeds the threshold, an incremental learning signal is immediately triggered. This signal is the driving identifier for starting full-link model optimization, used to drive closed-loop updates in manufacturing operation collaborative optimization, data model, knowledge graph, edge processing, and other links, enabling the system to correct decision logic and data model based on actual execution results, and achieve continuous self-optimization.

[0023] Furthermore, in the method provided in the application embodiments, edge processing is performed on the multidimensional manufacturing operation data to obtain a standardized manufacturing data stream. The method includes: The system locates edge gateway nodes for the Industrial Internet of Things (IIoT), performs outlier detection on multi-dimensional manufacturing operation data based on these nodes, and identifies anomalous outliers. It then cleans the multi-dimensional manufacturing operation data based on these outliers to obtain a cleaned dataset. Next, it performs data completion on the cleaned dataset using time-series interpolation to obtain a completed dataset. Finally, it performs normalization processing on the completed dataset, converts the normalization result to a format, and generates time-series data. The time-series data is then structured and encapsulated according to data type to generate the standardized manufacturing data stream.

[0024] Specifically, the first step is to perform the location operation of the industrial IoT edge gateway node. The edge gateway node is an edge computing hardware unit deployed between the sensing layer and the upper platform layer of the production site. It has the ability to convert multiple industrial protocols, perform local low-latency computing, and cache and forward data. It is a communication and computing bridge connecting the field sensing terminals and the upper management system. The location process is as follows: Based on the field network topology of the industrial IoT, all edge gateway nodes covering all production sensing nodes are identified, including industrial sensors, equipment programmable logic controllers (PLCs), RFID readers, smart meters, and vision acquisition terminals. The global unique identification encoding, network reachability verification, and real-time computing load assessment of each node are completed. Finally, the target edge gateway node that corresponds to the sensing terminals of the production line, workstation, and equipment that are the source of the multi-dimensional manufacturing operation data collected this time, has a direct communication link, and computing power redundancy that meets the preprocessing requirements is selected. At the same time, the communication link configuration between the sensing terminal and the target edge gateway node is completed, and the industrial transmission protocol, including Modbus, OPCUA, Profinet and other field common protocols, is adapted to ensure that the original multi-dimensional manufacturing operation data can be transmitted to the target edge gateway node with low latency and no packet loss.

[0025] After completing edge gateway node location and data transmission, outlier detection is performed on the received multidimensional manufacturing operation data based on the target edge gateway node to accurately identify abnormal outliers. This multidimensional manufacturing operation data comprises multi-source heterogeneous time-series data covering the entire intelligent manufacturing production chain, including four main categories: equipment status data characterizing the operating conditions of production equipment, process parameter data guiding production execution, material flow data recording the flow of production factors, and environmental monitoring data monitoring production site conditions. Abnormal outliers refer to invalid data that deviates from the statistical distribution range of data under normal production conditions, caused by factors such as sensor failure, network jitter, on-site electromagnetic interference, and instantaneous equipment start-up and shutdown disturbances. Failure to remove these outliers will directly lead to distortion in subsequent data modeling and analysis. Outlier detection is the operation of accurately locating and identifying this type of invalid data, specifically employing a two-stage approach: box plot statistical discrimination and isolated forest anomaly detection. The detection mechanism works as follows: The first level targets single-dimensional time-series data. It uses a box plot method to calculate the upper quartile (Q3) and lower quartile (Q1) of the data sequence, obtaining the interquartile range (IQR) = Q3 - Q1. The normal data range is set as [Q1 - 1.5IQR, Q3 + 1.5IQR]. Data exceeding this range is initially identified as suspected outliers. The second level targets multi-dimensional correlated data. It employs the isolated forest algorithm, using full-dimensional manufacturing operation data as input to construct multiple isolated trees. Anomaly scores are calculated based on the path length of data within the isolated trees. Data with anomaly scores exceeding a preset anomaly threshold are ultimately identified as outliers (default value 0.7, dynamically adjustable based on production conditions). Each identified outlier is labeled with its unique equipment identifier, data collection timestamp, data type, and anomaly type, achieving accurate location and traceable identification of all outliers.

[0026] After identifying outliers, data cleaning is performed on the multidimensional manufacturing operation data based on the labeled outliers to obtain a cleaned dataset. Data cleaning involves standardizing the identified outliers, filtering redundant, invalid, and incorrectly formatted data from the original data, eliminating noise interference, and ensuring the validity of the basic data. Specifically, for the labeled outliers, differentiated processing is performed based on the anomaly type. Single-transition outliers caused by transient electromagnetic interference or network packet loss are directly removed. Continuous outliers caused by short-term sensor failures or temporary equipment offline are handled differently. Outliers are first removed and their missing locations are marked to provide a precise basis for subsequent data completion. Then, redundant and invalid data are filtered on the dataset after outlier removal. Duplicate data collected from the same device, timestamp, and data type are deleted. Invalid data without valid device identifiers, timestamps, or formats that do not conform to industrial data standards are filtered out. At the same time, the units of measurement of all data are standardized and converted, such as temperature to degrees Celsius, pressure to megapascals, and rotation speed to r / min. Finally, a clean dataset with no obvious anomalies, a preliminary uniform format, consistent units of measurement, and no redundant or invalid content is obtained.

[0027] After data cleaning, the cleaned dataset is filled with missing data using time-series interpolation to obtain a complete dataset. Time-series interpolation is a method that addresses the temporal continuity of industrial manufacturing data by accurately filling in gaps left after outlier removal during data cleaning, as well as missing data due to network interruptions or temporary offline status of sensing nodes. Its purpose is to ensure the temporal continuity of manufacturing operation data. Specifically, this involves: first, accurately locating the missing data points in the cleaned dataset, based on a preset data collection frequency, such as 100ms / time or 1s / time. The process can be adjusted according to the type of sensing node. Based on the continuity of timestamps, all missing collection points and their corresponding data dimensions are identified. Then, differentiated temporal interpolation methods are used to distinguish the duration of the missing data: For short-term missing data with a duration ≤ 3 collection cycles, a linear interpolation method is used. Based on two adjacent valid data points before and after the missing time point, the completion value is calculated according to the linear proportion of the time interval. Specifically, the calculation is as follows: Let the valid time points before and after the missing time point t0 be t1 and t2, and the corresponding valid values ​​be v1 and v2. Then the completion value v0 = v1 + (v2 - v1) (t0-t1) / (t2-t1); For long-term missing data with a duration greater than 3 collection cycles, a time-series fitting interpolation method based on historical concurrent operating conditions is adopted. Historical normal production time-series data consistent with the current production conditions, i.e., the same product model, the same process standard, and the same equipment operating status, are retrieved. The least squares method is used to fit the time-series change pattern of this data dimension, and the completion value of the missing time point is calculated based on the fitting pattern. After completing the completion of all missing data, the time-series continuity of the completed dataset is checked to ensure that each collection cycle has corresponding valid data, without timestamp breaks or data dimension gaps. Finally, a completed dataset with continuous time series, complete dimensions, and no data gaps is obtained.

[0028] After completing the data missing data completion, normalization is performed on the completed dataset. The normalization results are then format-converted to generate time-series data. Normalization eliminates the dimensional differences between multi-source manufacturing data. Because different dimensions of the completed dataset come from different sensing terminals, their dimensions and value ranges vary significantly. For example, equipment speed ranges from 0-3000 r / min, ambient temperature from 10-40℃, and material inventory quantity from 0-1000 units. Directly using these dimensions for subsequent vector calculations and semantic association analysis would lead to the larger-range dimensions excessively dominating the calculation results. Therefore, normalization is necessary. All data is mapped to a unified numerical range, specifically using the Min-Max normalization method commonly used in industrial data processing: For each data dimension, the theoretical minimum value min_val and the theoretical maximum value max_val are first determined based on the historical valid data of that dimension under normal production conditions. Then, for each original value v of that dimension in the completed dataset, the formula v_norm=(v-min_val) / (max_val-min_val) is used to calculate and map all values ​​to the standard range of [0,1], thus completing the normalization process of all dimensions of data.

[0029] The format conversion process then involves transforming the scattered normalized data into time-series data with a unified timestamp identifier. This process integrates normalized data from all devices and all dimensions under the same collection timestamp, assigning each data set a globally unique timestamp identifier, a unique device identifier, and a data type code. The data is then sorted in ascending order according to chronological sequence, generating time-series data that is continuous in time dimension, accurately aligned across multiple dimensions, and has a unified format.

[0030] After the time series data is generated, it is structured and encapsulated according to data type to ultimately generate a standardized manufacturing data stream. This structured encapsulation involves encapsulating the time series data into fields and protocols according to a standardized data structure that can be uniformly recognized, read, and retrieved by the upper-level system. This ensures that subsequent semantic mapping and model building stages can directly retrieve the corresponding data. Specifically, the time series data is first divided into four standard data types based on its business attributes: equipment status, process parameters, material flow, and environmental monitoring. A unified structured field specification is set for each data type. Required fields include: globally unique data ID, equipment unique identifier code, data type code, acquisition timestamp, normalized value, original value, and unit of measurement. The process begins with identifying data source nodes and data quality labels. Following this, the time-series data is encapsulated row-by-row according to the specified field specifications. Simultaneously, the encapsulated data is converted to the JSON / OPCUA standardized data format commonly used in the Industrial Internet of Things (IIoT). The data is then streamed in chronological order to generate a standardized manufacturing data stream with a unified structure and protocol, directly readable by the upper-layer manufacturing semantic knowledge graph, and capable of stable streaming transmission. After encapsulation, the standardized manufacturing data stream undergoes data quality verification, including checks on temporal continuity, field completeness, and format compliance. Once verified, the data is transmitted to the upper-layer platform for subsequent semantic mapping and model building, and simultaneously cached locally on the edge gateway node for subsequent closed-loop updates and incremental learning.

[0031] Furthermore, in the method provided in the application embodiments, the standardized manufacturing data flow is mapped to a manufacturing semantic knowledge graph to construct a multi-level manufacturing operation data model. The method includes: Based on the standardized manufacturing data stream, equipment identification analysis is performed to extract multiple equipment identification tag data; based on the standardized manufacturing data stream, time feature analysis is performed to extract time series feature vectors; a manufacturing semantic knowledge graph is constructed, and semantic association calculation is performed between the time series feature vectors and the manufacturing semantic knowledge graph to construct a node-data association mapping matrix; according to the node-data association mapping matrix, the standardized manufacturing data stream is embedded into the manufacturing semantic knowledge graph according to the multiple equipment identification tag data to construct the multi-level manufacturing operation data model.

[0032] Specifically, the process begins with equipment identification analysis based on standardized manufacturing data streams. This involves extracting multiple equipment identification tags. The equipment identification analysis focuses on parsing, deduplicating, standardizing, and tagging the identity attribute data of production equipment bound to the standardized manufacturing data streams. Specifically, the structured fields of the standardized manufacturing data streams are parsed line by line to extract the unique equipment code, equipment model, production line number, workstation number, process compatibility range, and equipment communication address for each data entry. The extracted raw identification data is then deduplicated and verified for compliance, eliminating invalid data with no valid equipment identification or incomplete identification information. A globally unique and non-repeatable equipment identification code is generated for each physical production device. This identification code is then bound and encapsulated with all the corresponding identity attributes of the device, ultimately forming equipment identification tag data that corresponds one-to-one with each physical device. Each tag has a globally unique positioning capability, serving as a precise positioning anchor point for subsequent embedding of the standardized manufacturing data stream into the manufacturing semantic knowledge graph.

[0033] After simultaneously extracting equipment identification tag data, time feature analysis is conducted based on the standardized manufacturing data stream to extract time series feature vectors. This time feature analysis targets the inherent time series attributes of the standardized manufacturing data stream, mining the data's changing patterns, fluctuation characteristics, trends, and periodic properties along the time dimension. This transforms high-dimensional, discrete time-series numerical data into low-dimensional, dense, and computationally calculable standardized vectors. Specifically, based on the extracted equipment identification tag data, the standardized manufacturing data stream is grouped by equipment dimension and data type dimension, obtaining continuous time series data corresponding to each piece of equipment and each data dimension. For each group... Continuous time-series data is divided into time-series segments using a preset sliding window. The window size is dynamically set according to the data acquisition frequency, with a default setting of 10 consecutive acquisition cycles and a sliding step size of 1 acquisition cycle. Time-domain statistical features such as mean, variance, extreme values, rate of change, fluctuation period, and trend term are extracted for each time-series segment. Multiple sets of time-series segment features from the same device and of the same dimension are concatenated, and redundant features are removed and the feature dimensionality is reduced using the principal component analysis (PCA) algorithm. Finally, a time series feature vector with fixed dimensions, bound to the corresponding device identification label, and capable of completely representing the time dimension change pattern of the time-series data is generated, thus completing the featureization and vectorization transformation of the time-series data.

[0034] After simultaneously completing the extraction of two types of features, the construction of a manufacturing semantic knowledge graph is carried out. This graph is a semantic knowledge network used to structurally represent various business entities and explicit and implicit relationships between entities in the entire intelligent manufacturing scenario. It uses triples of head entity-relationship type-tail entity as its basic building blocks, transforming scattered industrial manufacturing knowledge into a computable, inferable, and associative structured knowledge system. After completing the construction of the manufacturing semantic knowledge graph, semantic association calculations are performed between the time-series feature vectors and the graph to construct a node-data association mapping matrix. This semantic association calculation uses deep learning algorithms and similarity matching algorithms to quantify the semantic relationships between the time-series feature vectors and the entity nodes in the manufacturing semantic knowledge graph. To establish a quantitative mapping relationship between time-series data streams and knowledge graph nodes, the following steps are taken: First, based on the constructed manufacturing semantic knowledge graph, the TransE graph embedding algorithm is used to extract the node embedding vector corresponding to each entity node in the graph. Then, the time-series feature vector bound to the device identification tag is used as the input sequence, and a bidirectional long short-term memory network (BiLSTM) is used to perform deep processing on the input sequence to generate a bidirectional hidden state vector that can simultaneously represent the bidirectional time dependency relationship between the time-series data. Next, the bidirectional hidden state vector is weighted and calculated through an attention mechanism to generate a semantic feature vector. Finally, semantic association mapping is carried out based on the semantic feature vector and the node embedding vector of the manufacturing semantic knowledge graph to construct a node-data association mapping matrix.

[0035] After constructing the node-data association mapping matrix, standardized manufacturing data flows are embedded into the manufacturing semantic knowledge graph according to multiple equipment identification tags, building a multi-layered manufacturing operation data model. Data embedding involves precisely binding the standardized manufacturing data flows to corresponding entity nodes in the knowledge graph based on association mapping relationships. Specifically, based on equipment identification tag data, the standardized manufacturing data flows are split by equipment dimension, ensuring each data flow corresponds to a unique equipment identification tag, and thus a unique physical production equipment. According to the node-data association mapping matrix, the corresponding equipment entity node in the manufacturing semantic knowledge graph for that equipment identification tag, as well as upstream and downstream related nodes with semantic associations to that data flow, including corresponding material entities, process entities, workstation entities, and production lines, are located. Entity nodes embed the standardized manufacturing data stream into corresponding entity nodes in real time according to timestamps and data types. These entity nodes serve as dynamic operational attribute data for the nodes and are deeply bound to the original static attribute data of the nodes, such as equipment models, rated parameters, and process standards. Ultimately, based on the inherent hierarchical entity structure of the manufacturing semantic knowledge graph, from the bottom-level entities of equipment, materials, and process parameters, to the middle-level entities of workstations and production lines, and then to the top-level entities of production plans and operation management, a full-level binding and semantic connection is formed from the bottom-level real-time operation data of equipment to the top-level operation management data. This ultimately constructs a multi-level manufacturing operation data model, which is a unified data model that covers all business levels of manufacturing, integrates static manufacturing knowledge and dynamic operational data, and has global semantic association and cross-dimensional reasoning capabilities.

[0036] Furthermore, in the method provided in the application embodiments, the process of constructing a semantic knowledge graph includes: Entity extraction is performed based on the multidimensional manufacturing operation data to construct a candidate entity set; relation extraction is performed based on the multidimensional manufacturing operation data to construct a candidate relation edge set; string similarity analysis is performed on the candidate entity set and the candidate relation edge set to calculate a first similarity; semantic similarity analysis is performed on the candidate entity set and the candidate relation edge set to calculate a second similarity; a weighted calculation is performed based on the first similarity and the second similarity, and the candidate entity set and the candidate relation edge set are merged according to the weighted result to obtain initial merged data; conflict analysis is performed based on the initial merged data to extract conflict relations, and the initial merged data is updated according to the conflict relations to determine the entity fusion set and the relation edge fusion set; association filling is performed based on the entity fusion set and the relation edge fusion set to construct an initial manufacturing semantic knowledge graph; the initial manufacturing semantic knowledge graph is quality evaluated, and when the quality evaluation meets the preset quality indicators, the manufacturing semantic knowledge graph is generated.

[0037] Specifically, the process begins with entity extraction based on multi-dimensional manufacturing operation data to construct a candidate entity set. Entity extraction involves identifying and extracting uniquely identifiable business entities with independent business meaning from multi-source heterogeneous manufacturing data throughout the intelligent manufacturing process. Specifically, for structured equipment nameplate parameter data, bill of materials (BOM) data, and process number data, the corresponding equipment model entity, material code entity, and process number entity are directly extracted. For unstructured process design documents and historical production log data, a named entity recognition model based on a combination of Bi-Short Memory Network (BiLSTM) and Conditional Random Field (CRF) is used to accurately extract equipment entity names, material entity names, process entity names, and parameter entity names. After completing the full entity extraction, all extracted entities are deduplicated and globally uniquely identified. Each entity is bound to its entity type, original data source, and attribute information, ultimately forming a candidate entity set that is non-repeating, traceable, and has standardized codes. This set covers all business entities to be verified throughout the intelligent manufacturing process.

[0038] After entity extraction is completed synchronously, relationship extraction is performed based on multi-dimensional manufacturing operation data to construct a candidate relationship edge set. Relationship extraction involves identifying and extracting explicit physical connections, implicit business associations, logical dependencies, and execution order constraints between different entities from multi-source manufacturing data. Specifically, it involves identifying the physical installation connections between equipment and the upstream and downstream serial relationships of the production process from the production line layout diagram data, extracting equipment-level relationship edges, and identifying the transfer paths, buffer locations, and flow sequences of materials between different workstations from the bill of materials (BOM) data and historical production log data, extracting material flow relationships. The process involves identifying the mutual constraints between different process parameters and the sequential dependencies of process execution from the process design document data, extracting process constraint relationship edges, and after completing the extraction of all relationship edges, standardizing all extracted relationship edges according to the standard triple format of industrial knowledge graph (head entity-relationship type-tail entity). Each relationship edge is bound with a globally unique identifier, relationship weight, original data source, and business scenario attributes, ultimately forming a set of candidate relationship edges that are non-repeating, traceable, and have a standardized format. This set covers all entity relationships to be verified in the entire intelligent manufacturing process.

[0039] After constructing the candidate entity set and candidate relation edge set, string similarity analysis is performed on both sets to calculate the first similarity. String similarity analysis is a quantitative analysis method that identifies synonymous entities and duplicate relations from the perspective of literal character overlap, targeting the name text of candidate entities and the description text of candidate relation edges. Specifically, the Levenstein edit distance algorithm, commonly used in industrial data processing, is used to calculate the edit distance for each pair of entity names in the candidate entity set, which is the minimum number of character addition, deletion, and modification operations required to convert one entity name string into another. At the same time, the edit distance for each pair of relation description texts in the candidate relation edge set is calculated. Then, the formula is used to quantify the calculation, obtaining the first similarity value between 0 and 1. The closer the value is to 1, the higher the literal overlap of the two comparison texts. Finally, the first similarity matrix of all entity pairs and relation edge pairs is formed.

[0040] After completing string similarity analysis, semantic similarity analysis is performed on the candidate entity set and the candidate relation edge set to calculate the second similarity. Semantic similarity analysis is a quantitative analysis method that identifies synonymous entities and relationships from a deep semantic level within industrial scenarios, targeting the business meaning of candidate entities and the business logic of candidate relation edges. For example, it identifies the synonymy of CNC machining center and CNC milling machine in machining scenarios, and material distribution and material transfer in logistics scenarios. Specifically, a pre-trained BERT language model from the industrial manufacturing domain is used to convert the name and attribute description text of each candidate entity and the type and description text of each candidate relation edge into fixed-dimensional semantic embedding vectors. Then, the cosine similarity formula is used to calculate the cosine value between the two vectors, obtaining a second similarity value ranging from 0 to 1. A value closer to 1 indicates a higher degree of industrial semantic overlap between the two compared texts, ultimately forming the second similarity matrix for all entity pairs and relation edge pairs. After completing the two types of similarity calculations, a weighted calculation is performed based on the first and second similarities. Based on the weighted results, the candidate entity set and the candidate relation edge set are aligned and merged to obtain an initial similarity matrix. The data merging process, where entity alignment and merging are based on similarity calculations, merges and unifies synonymous entities and duplicate relationships with literal or semantic consistency, eliminating redundant nodes and duplicate edges in the knowledge graph. Specifically, based on the business characteristics of the industrial manufacturing scenario, a preset string similarity weight α (default value 0.3) and semantic similarity weight β (default value 0.7) are used, with the sum of the weights being 1. The comprehensive similarity of each entity pair and relation edge pair is calculated using the formula: Total Similarity = α × First Similarity + β × Second Similarity. Simultaneously, a preset entity alignment threshold (default value 0.8) is used to further refine the data. Entity pairs with a similarity greater than or equal to the alignment threshold are identified as synonymous entities, and relation edge pairs are identified as duplicate relations. Synonymous entities are then merged, retaining a unique global entity ID and merging all attribute information, alias information, and source information of the two entities. Duplicate relation edges are then merged, retaining a unique relation edge ID and merging all attribute information, source information, and business scenario information of the relation edges. After merging all synonymous entities and duplicate relations, the initial merged data, which has undergone preliminary deduplication and normalization, is obtained. This data includes the preliminary merged entity set and relation edge set.

[0041] After obtaining the initial merged data, conflict analysis is conducted based on it to extract conflict relationships. The initial merged data is then updated according to these relationships, and the entity fusion set and relation edge fusion set are determined. Conflict analysis is crucial for identifying, verifying, and correcting logical contradictions and business conflicts in entity attributes and relational pointers within the initial merged data. Specifically, a rigid rule base for industrial manufacturing business is pre-built. This rule base includes rigid rules that conform to actual production practices, such as: the same physical equipment cannot simultaneously belong to two independent production lines with no business relationship; the process dependencies between the same operation cannot form a logical closed loop; the flow of production materials cannot involve meaningless circular backflows; and the rated parameters of the same entity cannot have two contradictory values. Based on this rule base, a full scan and verification of the initial merged data is performed to identify any data that does not conform to the business rules. The conflicts are categorized into three main types: entity attribute conflicts, relationship pointing conflicts, and business logic conflicts. The extracted conflicts are processed in a tiered manner. Minor conflicts, such as inconsistent units or non-standard attribute descriptions, are directly corrected according to industry standards. Severe conflicts, such as contradictory entity affiliations or logical errors in relationships, are verified and corrected by tracing back to the original data sources from entity and relationship extraction. After all conflicts are corrected, the initial merged data is updated synchronously, resulting in a conflict-free, logically inconsistent entity fusion set and a relationship edge fusion set that fully comply with industrial manufacturing business rules. The entity fusion set is the final knowledge graph node set after deduplication, merging, and conflict verification, while the relationship edge fusion set is the final knowledge graph edge set after deduplication, merging, and conflict verification. Together, they constitute the main body of the manufacturing semantic knowledge graph.

[0042] After determining the entity fusion set and the relation edge fusion set, association filling operations are performed based on the two sets to construct an initial manufacturing semantic knowledge graph. Association filling is based on existing entities and relation edges, using business logic reasoning to complete missing explicit relationships between entities and improve entity attribute links. Specifically: based on existing hierarchical relation edges, missing hierarchical relationships are completed through transitive logic reasoning. For example, if it is known that equipment A belongs to workstation 1, and workstation 1 belongs to production line X, the hierarchical relation edge of equipment A belonging to production line X is automatically completed. Based on process design documents and production log data, process parameter entities and corresponding equipment are completed. The binding relationships between entities and process entities are based on the Bill of Materials (BOM) data. The processing relationships between material entities and their corresponding processing process entities and production equipment entities are completed. After the full association is filled, the entities in the entity fusion set are used as nodes of the graph, and the relationship edges in the relationship edge fusion set are used as directed edges of the graph. Combined with the filled association relationships, an initial manufacturing semantic knowledge graph with a directed attribute graph structure that conforms to the industrial knowledge graph standard is constructed. This graph is the basic graph after the entity and relationship construction and preliminary association improvement are completed. It needs to be put into business application after quality assessment and iterative optimization.

[0043] After constructing the initial manufacturing semantic knowledge graph, a quality assessment is conducted. When the quality assessment meets preset quality indicators, the final manufacturing semantic knowledge graph is generated. The quality assessment quantifies the coverage, completeness, and richness of the initial manufacturing semantic knowledge graph. Specifically, it uses two quantitative indicators: graph density and relation completeness. Graph density characterizes the richness of relationships between entities in the graph, calculated as: Graph Density = Number of actual relation edges in the graph / Theoretically maximum number of relation edges that can be formed by all entities in the graph. Relationship completeness characterizes the completeness of the graph's coverage of the entire intelligent manufacturing process's business relationships, calculated as: Relationship Completeness = Number of covered business relationship types in the graph / Preset... The total number of business relationship types in intelligent manufacturing is set, along with preset quality indicators that meet business requirements, including a preset density threshold (default value 0.15, which can be dynamically adjusted according to business scenarios) and a preset integrity threshold (default value 0.9). When both the graph density index and the relationship integrity index are greater than or equal to the preset density threshold and the preset integrity threshold, the quality assessment is deemed passed, and the initial manufacturing semantic knowledge graph is determined as the final manufacturing semantic knowledge graph that can be used for business applications. If either indicator fails to meet the preset threshold, an incremental extraction process is triggered to supplement the missing entities and relationships from unprocessed historical production logs, process design documents, and production line operation data, iteratively updating the initial manufacturing semantic knowledge graph, and repeating the quality assessment process until both indicators meet the preset quality requirements.

[0044] Furthermore, in the method provided in the application embodiments, semantic association calculation is performed between the time series feature vector and the manufacturing semantic knowledge graph to construct a node-data association mapping matrix. The method includes: Multiple node embedding vectors are extracted based on the manufacturing semantic knowledge graph. The time-series feature vector is used as the input sequence, and a bidirectional long short-term memory (LSTM) network is employed to process the input sequence: S1: The bidirectional LSM network includes a forward LSM layer and a backward LSM layer; S2: The forward LSM layer processes the input sequence in forward chronological order, outputting a forward hidden state sequence; S3: The backward LSM layer processes the input sequence in reverse chronological order, outputting a backward hidden state sequence; The forward and backward hidden state sequences are concatenated to generate a bidirectional hidden state vector; An attention mechanism is used to calculate the bidirectional hidden state vector to generate a semantic feature vector; and a node-data association mapping matrix is ​​constructed based on the semantic feature vector and the manufacturing semantic knowledge graph through semantic association mapping.

[0045] Specifically, firstly, multiple node embedding vectors are extracted based on the constructed manufacturing semantic knowledge graph. These node embedding vectors are standardized vector representations obtained by mapping discrete entity nodes in the knowledge graph, which cannot be directly numerically calculated, to a low-dimensional dense vector space. This transforms the semantic attributes, association attributes, and business scenario attributes of entity nodes into vector forms that can be used for similarity matching and numerical calculation. Specifically, the TransE graph embedding algorithm, adapted to the triplet structure of industrial knowledge graphs, is used. For each entity node in the manufacturing semantic knowledge graph, including all types of business nodes such as equipment entities, material entities, process entities, workstation entities, and production line entities, vector training is conducted based on the triplet association relationships corresponding to that node. The node embedding vectors are generated with the same dimension as the time-series feature vectors of the preceding input. All node embedding vectors together form a set of node embedding vectors that correspond one-to-one with the entities in the graph, providing a unified computational foundation for subsequent semantic association matching. Then, the time-series feature vectors bound to the device identification tags are used as the input sequence. A bidirectional long short-term memory (BiLSTM) network is employed to perform deep temporal feature extraction on the input sequence. BiLSTM is a deep learning recurrent neural network specifically adapted for long-series data processing, capable of simultaneously capturing both forward and reverse time dependencies in time-series data. It overcomes the shortcomings of traditional unidirectional long short-term memory networks, which cannot capture inverse dependencies in time-series data and have insufficient ability to extract features from long sequences. Fully adaptable to the characteristics of long-cycle industrial manufacturing time-series data, strong dependencies between upstream and downstream processes, and strong correlations between upstream and downstream equipment, its processing strictly follows three standard steps: S1: Complete the standardized structure construction of a bidirectional long short-term memory network. This network includes parallel, independent, and structurally symmetrical forward long short-term memory layers and backward long short-term memory layers. The hidden layer dimensions and gating structures of the two network layers are completely identical, ensuring accurate matching of the output hidden state sequence dimensions and providing a compliant foundation for subsequent vector concatenation; S2: Process the input sequence in chronological order through the forward long short-term memory layer, that is, according to the chronological order of the time-series data collection time, input the time-series feature vectors sequentially from the initial collection time t0 to the latest collection time tn, and process them through the long short-term memory... The input gate, forget gate, and output gate of the memory network are standardized with gate control to filter out invalid temporal noise and retain key temporal features. The final output is a forward hidden state sequence containing the current information at each time step and information from all historical time steps. S3: The input sequence is processed in reverse chronological order through the backward long short-term memory layer. That is, the time series feature vector is input sequentially from the latest acquisition time tn to the initial acquisition time t0 in the reverse order of the time series data acquisition time. The temporal features are extracted through a gate control structure that is completely consistent with the forward layer. The final output is a backward hidden state sequence containing the current information at each time step and information from all future time steps. Through bidirectional temporal processing, the bidirectional dependencies between upstream and downstream processes and upstream and downstream equipment in industrial manufacturing time series data are fully captured.

[0046] After outputting the bidirectional hidden state sequence, the forward and backward hidden state sequences are concatenated to generate a bidirectional hidden state vector. The vector concatenation is a horizontal concatenation operation based on the vector dimension. Specifically, the forward and backward hidden state sub-vectors at the same acquisition time are merged by concatenating their first and last dimensions. For example, if the dimension of both the forward and backward hidden state sub-vectors is d, then the dimension of the concatenated single-time hidden state sub-vector is 2d. The concatenated sub-vectors at all times are integrated according to the original time order to finally generate a bidirectional hidden state vector that can simultaneously and completely represent the forward historical dependency and the reverse future dependency of the time series data, thus achieving the complete extraction of global time series features from the entire time series feature vector.

[0047] Subsequently, the bidirectional hidden state vectors are weighted using an attention mechanism to generate semantic feature vectors. Finally, semantic association mapping is performed based on the generated semantic feature vectors and the node embedding vectors used to create the semantic knowledge graph, constructing a node-data association mapping matrix. The semantic association mapping quantifies the semantic matching degree between the semantic feature vectors of time-series data and each entity node in the knowledge graph, establishing a quantitative association relationship between the time-series data stream and the knowledge graph nodes, and constructing a two-dimensional quantified node-data association mapping matrix. Each element value in this matrix accurately represents the semantic association degree between the corresponding time-series data feature and the corresponding knowledge graph entity node. Zero-value elements in the matrix represent that the corresponding node and the time-series feature have no effective semantic association, while non-zero-value elements represent that there is a clear semantic association between the two, and the larger the value, the higher the association degree.

[0048] Furthermore, in the method provided in the application embodiments, a semantic feature vector is generated by calculating the bidirectional hidden state vector through an attention mechanism. The method includes: The bidirectional hidden state vector is fully connected to perform the attention mechanism to obtain multiple original attention scores; the multiple original attention scores are normalized to obtain attention weights at multiple time steps; the bidirectional hidden state vector is matched and multiplied with the attention weights at multiple time steps to generate a weighted result at multiple time steps; the weighted result at multiple time steps is added element-wise according to the bidirectional hidden state vector to generate the semantic feature vector.

[0049] Specifically, firstly, a fully connected computation is performed on the bidirectional hidden state vector using an attention mechanism to obtain multiple raw attention scores. The attention mechanism is a weight allocation algorithm specifically adapted to the characteristics of long-term data in industrial manufacturing, capable of automatically identifying and amplifying high-business-value time-series features while suppressing invalid noise features. Here, a single-hidden-layer fully connected neural network structure is used for attention layering. Its weight parameters are pre-trained and iteratively optimized using labeled datasets of historical normal and abnormal production conditions. The specific process of the fully connected computation is as follows: the independent sub-vectors corresponding to each time step in the bidirectional hidden state vector are sequentially input into the attention layered fully connected network. Linear transformation and nonlinear activation are performed through trainable weight matrices and bias terms, mapping each high-dimensional time-step sub-vector to a one-dimensional scalar value. This value is the raw attention score for the corresponding time step. The number of raw attention scores is completely consistent with the number of time steps of the bidirectional hidden state vector, and their magnitude initially represents the contribution of the time-series features at the corresponding time step to the final manufacturing business semantics.

[0050] Subsequently, the multiple raw attention scores were normalized to obtain attention weights at multiple time points. The normalization process was implemented using the Softmax activation function, which is commonly used for weight allocation in industrial time-series data. The specific calculation process is as follows: all raw attention scores are subjected to an exponential transformation, and the result of the exponential transformation at each time point is divided by the sum of the exponential transformation results at all time points. Finally, the raw attention scores, which originally had no constraints on their value range, are transformed into standardized weight values ​​with fixed values ​​between 0 and 1, and the sum of the corresponding values ​​at all time points is strictly equal to 1. This weight value is the attention weight at the corresponding time point. Each attention weight corresponds one-to-one with the collection time point, and its value accurately quantifies the contribution ratio of the temporal features at the corresponding time point to the final semantic feature construction. The higher the value, the higher the value of the manufacturing data features at that time point in representing the production and operation status. For example, critical time points such as abnormal equipment fluctuations, process parameter exceedances, and production process switching will correspond to higher attention weights, while normal time points with stable equipment operation and no parameter fluctuations will correspond to lower attention weights.

[0051] Next, the bidirectional hidden state vector is matched and multiplied with the attention weights at multiple time points to generate a weighted result for multiple time points. The matching multiplication strictly follows the one-to-one correspondence between the acquisition time points. Each sub-vector in the bidirectional hidden state vector at each time point is multiplied by the scalar attention weight corresponding to that time point. That is, each dimension element in the sub-vector is multiplied by the corresponding attention weight, thereby realizing the weight allocation of the temporal features at each time point. The amplitude of high-value key features is amplified, while the amplitude of low-value noise features and conventional stationary features is suppressed. Finally, a weighted result for multiple time points is generated with the same dimensions as the original bidirectional hidden state vector, with each time point corresponding to a weighted sub-vector. This completes the differentiated weight adaptation of the full temporal features.

[0052] Finally, the weighted results from multiple time points are summed element-wise according to the bidirectional hidden state vector to generate a semantic feature vector. The element-wise summation is performed by summing the elements of the weighted sub-vectors from all time points at the same dimension, while strictly maintaining the alignment of the vector dimensions. The result is a fixed-dimensional dense vector with the same dimension as the single-time bidirectional hidden state sub-vector. This vector is the semantic feature vector, which fully integrates the business semantics of the manufacturing time-series data throughout the entire acquisition cycle. It filters out invalid noise caused by on-site electromagnetic interference and network jitter, and highlights key features with strong characterization capabilities for production operation status, equipment health, process execution effect, and material flow efficiency. It can be directly used for subsequent semantic association matching calculations with the manufacturing semantic knowledge graph.

[0053] Furthermore, in the method provided in the application embodiments, a node-data association mapping matrix is ​​constructed based on the semantic feature vector and the manufacturing semantic knowledge graph. The method includes: Based on the semantic feature vector and the manufacturing semantic knowledge graph, node matching is performed to extract the embedding vectors of multiple nodes; cosine similarity is calculated based on the embedding vectors of the multiple nodes to obtain multiple cosine similarity values; a similarity threshold is set, and the multiple cosine similarity values ​​are compared with the similarity threshold; target nodes with cosine similarity values ​​greater than the similarity threshold are extracted, and the target nodes are determined as the first nodes with semantic association with the time series feature vector. The above process is repeated until all nodes are traversed to obtain multiple associated nodes; the multiple associated nodes are matched and integrated with the multiple cosine similarity values ​​to construct the node-data association mapping matrix.

[0054] Specifically, firstly, node matching is performed based on semantic feature vectors and the manufacturing semantic knowledge graph to extract embedding vectors for multiple nodes. Node matching is based on equipment identifier tags bound to semantic feature vectors. First, the corresponding equipment entity node in the manufacturing semantic knowledge graph is located. Then, based on the association links of the graph, the related entity nodes such as materials, processes, workstations, and production lines upstream and downstream of the node are covered. Finally, it is expanded to all business entity nodes in the graph to complete the full coverage of the nodes to be matched. Subsequently, the TransE graph embedding pre-trained model adapted to the industrial knowledge graph is used to extract the node embedding vectors corresponding to all nodes to be matched. The node embedding vector is a standardized vector representation obtained by mapping the unstructured information such as static attributes, associations, and business semantics of entity nodes in the graph to a low-dimensional dense vector space. Its vector dimension is strictly consistent with the dimension of the input semantic feature vector, providing a unified numerical calculation basis for subsequent similarity calculation. Each node embedding vector is bound to the corresponding entity node and has a globally unique identifier.

[0055] After extracting the embedding vectors of all nodes to be matched, cosine similarity calculation is performed based on the embedding vectors of multiple nodes to obtain multiple cosine similarity values. Cosine similarity is an indicator used in industrial semantic matching scenarios to quantify the degree of directional overlap between two vectors in vector space, thereby representing the degree of semantic association between them. Its standardized calculation formula is: Cosine similarity = (dot product of node embedding vector and semantic feature vector) / (magnitude of node embedding vector × magnitude of semantic feature vector). Using this formula, the embedding vector of each node is calculated with the input semantic feature vector one by one to obtain the cosine similarity value corresponding to each node, which is strictly between -1 and 1. The closer the value is to 1, the higher the overlap between the business semantics of the node and the semantic feature vector representing the manufacturing operation data semantics. The closer the value is to 0, the less effective the semantic association between the two. All calculated cosine similarity values ​​are bound to the corresponding entity nodes one by one to form a traceable matching result set.

[0056] After calculating all cosine similarity values, a similarity threshold is set. Multiple cosine similarity values ​​are compared one by one with this threshold. The similarity threshold is a pre-defined threshold based on business verification data from an industrial manufacturing scenario, used to determine whether there is a valid business semantic association between entity nodes and semantic feature vectors. The default value is 0.6, which can be dynamically adjusted according to the needs of specific business scenarios. For example, for high-precision scenarios involving equipment anomaly handling, the threshold can be increased to 0.75; for high-recall scenarios involving full production line status awareness, the threshold can be decreased to 0.5. After setting the threshold, the cosine similarity value corresponding to each node is compared one-to-one with the threshold, and the magnitude of each comparison result is marked, providing a basis for subsequent filtering of associated nodes. After completing the threshold comparison, the cosine similarity values ​​greater than the similarity threshold are extracted. The target node is identified as the first node with a semantic association with the time series feature vector. Based on the above process, the determination is repeated until all nodes are traversed, resulting in multiple associated nodes. The first node is the first entity node that passes the threshold determination and has a valid semantic association with the semantic feature vector. The repeated determination refers to repeatedly executing the entire process of node embedding vector extraction, cosine similarity calculation, and threshold comparison. All entity nodes in the manufacturing semantic knowledge graph that have not been matched are checked one by one to ensure that no node is missed. Finally, all entity nodes with cosine similarity values ​​greater than the preset similarity threshold are selected to form multiple associated nodes. Each associated node is bound to a corresponding unique node identifier, node type, and the matched cosine similarity value, fully covering all business entities with semantic association with the current manufacturing operation data.

[0057] After completing the full screening of associated nodes, a node-data association mapping matrix is ​​constructed by matching and integrating multiple associated nodes with their corresponding cosine similarity values. This matching and integration involves precisely binding each associated node to its corresponding cosine similarity value, while simultaneously defining the positioning rules for each associated node within the matrix. The node-data association mapping matrix is ​​a two-dimensional standardized matrix used to quantify the degree of semantic association between standardized manufacturing data streams and entity nodes in the manufacturing semantic knowledge graph. Its specific construction rules are as follows: the entire set of entity nodes in the manufacturing semantic knowledge graph serves as the row dimension of the matrix, with each row corresponding to a uniquely identified entity node. The semantic feature vector corresponds to the feature dimensions of the standardized manufacturing data flow, including four dimensions: equipment status, process parameters, material flow, and environmental monitoring. These are the column dimensions of the matrix, with each column corresponding to a type of manufacturing data feature. The element value at each position in the matrix is ​​the cosine similarity value between the entity node in the corresponding row and the manufacturing data feature in the corresponding column. For entity nodes that are not identified as associated nodes, the corresponding matrix element value is uniformly set to 0. For entity nodes that are identified as associated nodes, the corresponding matrix element value is filled with the cosine similarity value obtained from the matching. Finally, a node-data association mapping matrix with a standardized structure is formed, which can be directly used for subsequent data embedding and business reasoning.

[0058] Furthermore, in the method provided in the application embodiments, manufacturing operation collaborative optimization is performed based on the multi-level manufacturing operation data model to formulate an intelligent decision instruction set. The method includes: Based on the multi-level manufacturing operation data model, manufacturing operation collaborative optimization analysis is performed to construct manufacturing operation collaborative optimization process data. This process data is decomposed into task scheduling optimization sub-branches, resource allocation optimization sub-branches, and exception handling optimization sub-branches, which are parallel processes. The task scheduling optimization sub-branches read data from the multi-level manufacturing operation data model to obtain production plan data and equipment availability data. Based on the production plan data and equipment availability data, mixed integer programming is performed to generate task priority sequences and equipment allocation schemes, which are then encapsulated to construct task scheduling decision instructions. Finally, the resource allocation optimization sub-branches are used to optimize the multi-level manufacturing operation data model. The model reads data to obtain material inventory data, workstation load data, and logistics path data. Based on the material inventory data, workstation load data, and logistics path data, multi-objective optimization is performed to generate material delivery paths and workstation load balancing schemes, which are then encapsulated to construct resource allocation decision instructions. The anomaly handling optimization sub-branch reads data from the multi-level manufacturing operation data model to obtain equipment status anomaly data and process parameter over-limit data. Based on the equipment status anomaly data and process parameter over-limit data, hybrid reasoning is performed to generate anomaly handling schemes and early warning notification data, which are then encapsulated to construct anomaly handling decision instructions. The task scheduling decision instructions, resource allocation decision instructions, and anomaly handling decision instructions are merged and sorted to generate the intelligent decision instruction set.

[0059] Specifically, the first step is to conduct collaborative optimization analysis of manufacturing operations based on this multi-level manufacturing operations data model. This analysis focuses on the business objectives of intelligent manufacturing operations, including maximizing on-time order delivery rate, maximizing overall equipment utilization rate, minimizing production and operating costs, and maximizing the closed-loop response speed to production anomalies. By combining the associated data of all business entities in the data model, the entire business process from order placement, production scheduling, material distribution, process processing to finished product warehousing is analyzed. The optimization objectives, rigid constraint boundaries, data input sources, and cross-process business linkage rules for each business link are clarified, ultimately constructing the collaborative optimization process data for manufacturing operations. This involves the digital decomposition and standardized path definition of the entire business optimization chain in intelligent manufacturing operations. This standardizes the execution boundaries and collaborative logic of each subsequent optimization branch, fundamentally preventing issues such as target conflicts, resource contention, and execution sequence disorder among optimization modules. Subsequently, the manufacturing operation collaborative optimization process data is decomposed into three fully parallel optimization branches: task scheduling optimization sub-branch, resource allocation optimization sub-branch, and anomaly handling optimization sub-branch. It is clarified that the three branches synchronously retrieve the same source business data based on the same multi-level manufacturing operation data model to carry out parallel optimization calculations, ensuring data consistency, timing synchronization, and execution efficiency in the optimization process, and avoiding the decision lag and collaboration deviation problems caused by the traditional serial optimization mode.

[0060] The optimization calculations of three parallel branches are initiated simultaneously. First, the task scheduling optimization sub-branch is based on the semantic association link of the multi-level manufacturing operation data model, synchronously reading the full amount of production plan data and equipment availability data. The production plan data is a guiding data for production execution, containing information such as order delivery cycle, product BOM structure, process sequence dependencies, target output, and quality control standards. It has been synchronized to the corresponding node of the operation layer of the data model through semantic mapping. The equipment availability data is generated based on real-time equipment status data, historical maintenance records, equipment rated processing parameters, and process adaptability range. It can accurately characterize the processing time, rated capacity, availability status, and changeover cost of each production equipment within the planning cycle. Subsequently, based on the above two types of data, the optimization objectives are to minimize order delivery delays and maximize overall equipment utilization rate. The rigid constraints are the process execution order, equipment processing adaptability, planning cycle boundary, and equipment capacity limit. A mixed integer approach is used. The planning algorithm performs global optimization, among which mixed integer programming is a mathematical optimization algorithm specifically adapted to production task scheduling scenarios where process tasks and equipment allocation are discrete integer variables and constraints are linear business rules. It can obtain the global optimal solution under the premise of satisfying all rigid constraints, and finally generate a standardized task priority sequence and equipment allocation scheme. The task priority sequence is an execution sequence that prioritizes all production process tasks to be executed according to multiple dimensions such as order urgency, process dependencies, equipment processing efficiency, and quality control requirements, and clarifies the execution order and time limit requirements of each task. The equipment allocation scheme is an execution plan that matches the optimal available processing equipment for each process task, clarifies the task processing sequence of a single piece of equipment, production changeover arrangements, and capacity reservation. The above task priority sequence and equipment allocation scheme are then encapsulated in a standardized format that can be recognized by the underlying execution system of the industrial site to generate task scheduling decision instructions that can be directly issued and executed.

[0061] The resource allocation optimization sub-branch, which executes in parallel with the task scheduling optimization sub-branch, synchronously reads all material inventory data, workstation load data, and logistics path data based on the semantic association links of the multi-level manufacturing operation data model. The material inventory data includes real-time inventory quantities, storage locations, specifications, applicable processes, and completeness status of raw materials, work-in-process, and semi-finished products, fully covering the entire inventory node status across the central warehouse, line-side warehouse, and workstation buffer area. The workstation load data is calculated based on the amount of pending tasks, rated processing time, occupied working hours, and personnel configuration status of each workstation. The obtained data accurately characterizes the real-time load rate, waiting queue duration, and capacity redundancy of each workstation, directly reflecting the load balance status and production bottleneck locations of each workstation across the entire production line. Logistics path data is constructed based on the workshop's digital layout, production line workstation locations, material handling equipment access rules, and historical logistics time and congestion data. This standardized road network data covers the entire workshop area, including feasible logistics paths, path lengths, rated passage times, and congestion probabilities. Subsequently, based on these three types of data, the optimization objectives are multi-dimensional and mutually constraining: highest material delivery on-time rate, optimal workstation load balance, and lowest overall logistics cost. With rigid constraints such as production cycle time requirements, material availability time limits, maximum workstation load thresholds, and logistics equipment access rules, a non-dominated sorting genetic algorithm (NSGA-II) with an elitist strategy is employed to solve multi-objective optimization problems. Multi-objective optimization addresses multiple mutually constraining optimization objectives in resource allocation scenarios, providing a Pareto optimal solution set that balances multiple objectives. This effectively avoids the problem of single-objective optimization causing other objectives to deviate significantly from their optimal values. Ultimately, standardized material delivery routes and workstation load balancing schemes are generated, where the material delivery route is planned for each material delivery task. The optimal material handling path, delivery time, delivery batch, and delivery equipment allocation scheme can ensure that materials are supplied to production workstations on time and in place, while avoiding excessive backlog of line-side inventory. The workstation load balancing scheme is based on the real-time load data of each workstation and dynamically allocates process tasks across workstations. Redundant tasks of high-load workstations are allocated to low-load idle workstations with capacity redundancy, thus balancing the workstation load of the entire production line and eliminating production bottlenecks. The above material delivery path and workstation load balancing scheme are then packaged in a standardized format that can be recognized by the industrial site logistics execution system to generate resource allocation decision instructions that can be directly issued and executed.

[0062] The anomaly handling optimization sub-branch, executed in parallel with the first two branches, synchronously reads all equipment status anomaly data and process parameter exceedance data based on the semantic association link of the multi-level manufacturing operation data model. Equipment status anomaly data is identified by comparing real-time collected equipment operating data with preset equipment health standards, identifying anomalies such as equipment failure, decreased accuracy, operating parameters exceeding thresholds, and maintenance warnings. This data includes the unique identifier of the abnormal equipment, anomaly type, anomaly level, anomaly location, and the scope of its impact on production. Process parameter exceedance data is identified by comparing real-time collected field process parameters with preset process standard ranges, identifying process anomalies such as parameter exceedances, parameter drift, and process execution deviations. This data includes the abnormal workstation, abnormal parameter type, exceedance range, and impact level on product quality. Subsequently, based on these two types of anomaly data, a hybrid reasoning mechanism is used to optimize the anomaly handling process. This hybrid reasoning combines rule-based deterministic reasoning with case-based reasoning. The example uses a composite reasoning mechanism based on fuzzy reasoning. First, deterministic rule reasoning is performed based on a pre-set equipment failure handling rule library and process anomaly handling specification library to match the standard handling process for the corresponding anomaly. Then, combined with a library of successful handling cases of similar historical anomalies, case reasoning is performed through semantic similarity matching to supplement and optimize the standard handling process, making it fully adaptable to the specific production scenario of the current anomaly. Finally, standardized anomaly handling plans and early warning notification data are generated. The anomaly handling plan is formulated for specific anomaly events and includes a feasible plan that includes the person in charge of handling, handling steps, required spare parts and materials, estimated downtime, and suggestions for temporary adjustments to the production process. The early warning notification data is standardized early warning information that is pushed to the corresponding management personnel, maintenance personnel, and production team leaders according to the anomaly level. It includes anomaly information, handling requirements, and response time limits. Finally, the above anomaly handling plan and early warning notification data are packaged in a standardized format that can be recognized by the industrial site operation and maintenance management system to generate anomaly handling decision instructions that can be directly issued and executed.

[0063] After the three parallel optimization branches have completed the generation of their respective decision instructions, the task scheduling decision instructions, resource allocation decision instructions, and anomaly handling decision instructions are subjected to collaborative verification and merging sorting. Collaborative verification involves adjusting instructions with execution conflicts, resource contention, or timing contradictions to ensure that all instructions have consistent execution goals, conflict-free resource allocation, and self-consistent timing logic. Merging sorting prioritizes instructions according to their urgency level, execution timing dependency, and scope of impact. High-level anomaly handling instructions take precedence over regular task scheduling and resource allocation instructions, and task scheduling instructions for preceding processes take precedence over instructions for subsequent processes. Ultimately, a unified, collaborative, conflict-free, and closed-loop executable intelligent decision instruction set is formed. This intelligent decision instruction set is a standardized set of executable instructions covering all scenarios of production task scheduling, dynamic resource allocation, and production anomaly handling. It serves as the output carrier for the entire manufacturing operation management solution, from data perception and semantic modeling to implementation.

[0064] Furthermore, in the method provided in the application embodiments, the execution of the intelligent decision instruction set involves real-time monitoring of execution feedback data, and closed-loop updates are performed based on the execution feedback data. The method includes: The intelligent decision-making instruction set is sent to the equipment controller and material handling system to trigger equipment control instructions and material handling instructions. Real-time data acquisition is performed on the equipment control instructions and material handling instructions to obtain actual equipment status change data, actual material flow data, and actual production cycle data. These data are used as execution feedback data. Execution expectation analysis is performed based on the intelligent decision-making instruction set to set expected execution goals. The execution feedback data and the intelligent decision-making instruction set are compared and analyzed item by item according to the expected execution goals to calculate multiple deviation components. These multiple deviation components are synthesized according to data direction to construct a deviation vector. The length of the deviation vector is calculated to obtain a comprehensive length value, which is then converted to generate a comprehensive deviation metric. When the comprehensive deviation metric exceeds a preset deviation threshold, an incremental learning signal is triggered to perform a closed-loop update for manufacturing operation collaborative optimization.

[0065] Specifically, the intelligent decision-making instruction set is first transmitted through the industrial IoT communication network established in the early stages of the solution. The edge gateway nodes located in the preceding steps then perform conversion and precise distribution of multiple industrial protocols such as Modbus, OPCUA, and Profinet. These instructions are then sent to the corresponding equipment controllers and material handling systems on the production floor. The equipment controllers refer to the programmable logic controllers (PLCs) and computer numerical control systems (CNCs) that power various processing and testing equipment on the production floor. They can receive and parse standardized control instructions, driving the equipment to complete corresponding processing actions, parameter adjustments, start-stop control, and other operations. The material handling system refers to the in-plant logistics execution system composed of AGV / RGV transport vehicles, automated storage and retrieval systems, roller conveyors, and line-side warehouse management units. It can receive and execute instructions for material entry / exit, transfer, distribution, and buffer management. After the instructions are sent, the corresponding equipment control instructions and material handling instructions are triggered simultaneously, officially initiating the entire production process.

[0066] Throughout the entire lifecycle execution of the intelligent decision-making instruction set, the industrial IoT sensing and acquisition system built in the initial stage of the solution uses a high-frequency synchronous acquisition frequency completely consistent with the initial multi-dimensional manufacturing operation data acquisition to collect real-time data on the execution status of the production site in all time periods and dimensions. This acquires actual equipment status change data, actual material flow data, and actual production cycle data that directly correspond to the effect of instruction execution. The actual equipment status change data refers to the real-time operating parameters, start / stop status, processing progress, utilization rate, accuracy status, and fault alarm information of the equipment after the execution of equipment control instructions. This directly reflects the equipment's execution and matching of scheduling instructions. The actual material flow data refers to the actual entry and exit times and on-the-go locations of materials after the execution of material handling instructions. Actual logistics data, such as delivery arrival time, line-side warehouse inventory dynamics, and process completeness status, directly reflect the execution and implementation effect of material delivery instructions. Actual production cycle data refers to the actual processing time of each process, the overall production line operating cycle, order completion progress, inter-process transfer time, and product qualification rate after the execution of the entire process instructions. This data directly reflects the achievement of the overall decision instructions on the production plan objectives. The three types of data collected undergo a completely consistent edge processing process with the initial stage of the solution, completing outlier detection, data cleaning, missing data completion, normalization, and structured encapsulation to generate a standardized feedback data stream. Finally, these three types of standardized data are unified as execution feedback data, providing a high-quality, formatted, and time-aligned analytical foundation for subsequent deviation analysis.

[0067] After standardizing the execution feedback data, execution expectation analysis is conducted based on the issued intelligent decision-making instruction set. This analysis involves breaking down each sub-instruction in the intelligent decision-making instruction set into quantifiable and comparable sub-execution benchmarks according to its corresponding optimization goals and business management requirements. This clarifies the expected completion time, rated execution parameters, target achievement value, and constraint boundaries for each instruction, ultimately forming expected execution targets that correspond one-to-one with the instruction set. This standardized benchmark system is used to evaluate the execution effectiveness of the instructions. Subsequently, the standardized execution feedback data is compared with... The corresponding expected execution targets are compared item by item, and the quantitative difference between the actual value and the benchmark value is calculated for each comparison dimension, generating multiple deviation components. Each deviation component corresponds to the execution deviation of a single business dimension, such as process progress deviation component, material delivery on-time rate deviation component, equipment utilization rate deviation component, production cycle time deviation component, process parameter execution deviation component, etc. Each deviation component is clearly marked with the deviation direction and magnitude, which can accurately locate the execution deviation of a single dimension. Then, the multiple deviation components with directional attributes are synthesized according to the preset business dimension order, first combining multiple deviation components of different dimensions and units. Following the same minimum-maximum normalization process as the previous steps in the scheme, and mapping to a unified numerical range of [0,1], a deviation vector is constructed. Each element of the deviation vector corresponds to a normalized deviation component. The direction of the vector represents the source dimension of the deviation, and the magnitude of the vector represents the severity of the overall deviation, thus achieving a unified vectorized representation of multi-dimensional dispersed deviations. After constructing the deviation vector, its length is calculated using the L2 norm method, which involves taking the square root of the sum of the squares of all elements in the deviation vector to obtain the comprehensive length value. This comprehensive length value is then normalized to generate a value. The deviation measurement value is strictly within the range of [0,1]. The closer the value is to 1, the greater the overall deviation between the actual execution effect and the decision expectation, and the worse the execution effect. At the same time, a preset deviation threshold is set, which can be dynamically adjusted according to production management requirements. This threshold is the maximum tolerance limit of production execution. The default setting is 0.2, which can be adjusted up or down according to the urgency of the order, the accuracy requirements of the production process, and the quality control standards. The deviation measurement value is compared with the preset deviation threshold. When the deviation measurement value does not exceed the preset deviation threshold, it is determined that the execution of this instruction meets the expectations, and only the feedback data of this execution is stored in the historical production database.

[0068] When the comprehensive deviation metric exceeds the preset deviation threshold, it is determined that the execution of this instruction has seriously deviated from expectations, and the existing decision optimization model can no longer adapt to the current production conditions. An incremental learning signal is immediately triggered. This incremental learning signal is a closed-loop update trigger signal carrying the deviation source dimension, deviation quantification data, corresponding execution feedback data, and real-time production condition data. It is used to initiate a closed-loop optimization update across the entire solution chain. The specific update process strictly revolves around the entire manufacturing operations collaborative optimization process, including locating the source of the current execution deviation based on the direction and component weights of the deviation vector, and optimizing the mixed-integer programming model of the task scheduling optimization sub-branch. The algorithm includes a rule base and a case base for the hybrid reasoning mechanism of the sub-branch for optimizing resource allocation and the weight of the target, as well as the adaptation parameters of the multi-objective optimization algorithm. The incremental data is also used to optimize the multi-level manufacturing operation data model, the semantic association matching logic of the manufacturing semantic knowledge graph, and the algorithm parameters for edge data processing in the preceding stages. After the full-link update is completed, all optimized models, algorithms, and rules are deployed to the corresponding execution units and applied to the next round of intelligent manufacturing operation management, realizing closed-loop control of the entire lifecycle from decision-making to execution, monitoring, analysis, optimization, and re-decision-making.

[0069] In summary, the intelligent manufacturing operation management method based on the Industrial Internet of Things provided in this application has the following technical effects: Through a comprehensive technical solution encompassing edge data standardization processing, manufacturing semantic knowledge graph construction, multi-level operational data model mapping, multi-branch parallel collaborative optimization, and closed-loop execution updates, high-quality standardized transformation of multi-source manufacturing operation data has been achieved. This solution breaks down the semantic barriers between data and production business entities, enabling parallel collaboration and conflict-free decision output in task scheduling, resource allocation, and anomaly handling. It also establishes a quantitative monitoring system for decision execution deviations and a closed-loop self-optimization mechanism across the entire chain. This effectively improves the accuracy of production decisions and the speed of on-site execution response, increases the overall equipment utilization rate and the efficiency of production resource utilization, and achieves efficient, collaborative, and intelligent control of intelligent manufacturing operations management, thereby reducing operating costs.

[0070] Example 2, based on the same inventive concept as the intelligent manufacturing operation management method based on the Industrial Internet of Things in the foregoing examples, such as... Figure 2 As shown in the embodiment of this application, a smart manufacturing operation management system based on the Industrial Internet of Things is provided. The system includes: The processing module 11 is used to collect multi-dimensional manufacturing operation data through industrial IoT sensing, perform edge processing on the multi-dimensional manufacturing operation data, and obtain a standardized manufacturing data stream; the mapping module 12 is used to map the standardized manufacturing data stream to a manufacturing semantic knowledge graph to construct a multi-level manufacturing operation data model; the optimization module 13 is used to perform collaborative optimization of manufacturing operations based on the multi-level manufacturing operation data model and formulate an intelligent decision instruction set; the execution module 14 is used to execute the intelligent decision instruction set, monitor the execution feedback data in real time, and perform closed-loop updates based on the execution feedback data.

[0071] Furthermore, the processing module 11 is also used to perform the following steps: locating the edge gateway node of the industrial Internet of Things; performing outlier detection on the multidimensional manufacturing operation data based on the edge gateway node to identify abnormal outliers; cleaning the multidimensional manufacturing operation data based on the abnormal outliers to obtain a cleaned dataset; performing data missing completion on the cleaned dataset according to time-series interpolation to obtain a completed dataset; performing normalization processing on the completed dataset; converting the normalization processing result into a format to generate time-series data; and structurally encapsulating the time-series data according to data type to generate the standardized manufacturing data stream.

[0072] Furthermore, the mapping module 12 is also used to perform the following steps: perform equipment identification analysis based on the standardized manufacturing data stream to extract multiple equipment identification tag data; perform time feature analysis based on the standardized manufacturing data stream to extract time series feature vectors; construct a manufacturing semantic knowledge graph, perform semantic association calculation between the time series feature vectors and the manufacturing semantic knowledge graph, and construct a node-data association mapping matrix; embed the standardized manufacturing data stream into the manufacturing semantic knowledge graph according to the multiple equipment identification tag data according to the node-data association mapping matrix, and construct the multi-level manufacturing operation data model.

[0073] Furthermore, the mapping module 12 is also used to perform the following steps: extracting entities based on the multidimensional manufacturing operation data to construct a candidate entity set; extracting relationships based on the multidimensional manufacturing operation data to construct a candidate relationship edge set; performing string similarity analysis on the candidate entity set and the candidate relationship edge set to calculate a first similarity; performing semantic similarity analysis on the candidate entity set and the candidate relationship edge set to calculate a second similarity; performing weighted calculation based on the first similarity and the second similarity, and merging the candidate entity set and the candidate relationship edge set according to the weighted result to obtain initial merged data; performing conflict analysis based on the initial merged data to extract conflict relationships, updating the initial merged data according to the conflict relationships, and determining the entity fusion set and the relationship edge fusion set; performing association filling based on the entity fusion set and the relationship edge fusion set to construct an initial manufacturing semantic knowledge graph; and performing quality evaluation on the initial manufacturing semantic knowledge graph, generating the manufacturing semantic knowledge graph when the quality evaluation meets the preset quality indicators.

[0074] Furthermore, the mapping module 12 is also used to perform the following steps: extracting multiple node embedding vectors based on the manufacturing semantic knowledge graph; using the time series feature vector as the input sequence, and processing the input sequence using a bidirectional long short-term memory network: S1: the bidirectional long short-term memory network includes a forward long short-term memory layer and a backward long short-term memory layer; S2: processing the input sequence in forward time order through the forward long short-term memory layer to output a forward hidden state sequence; S3: processing the input sequence in reverse time order through the backward long short-term memory layer to output a backward hidden state sequence; concatenating the forward hidden state sequence and the backward hidden state sequence to generate a bidirectional hidden state vector; calculating the bidirectional hidden state vector through an attention mechanism to generate a semantic feature vector; and constructing a node-data association mapping matrix based on the semantic feature vector and the manufacturing semantic knowledge graph through semantic association mapping.

[0075] Furthermore, the mapping module 12 is also used to perform the following steps: performing fully connected computation on the bidirectional hidden state vector through the attention mechanism to obtain multiple original attention scores; normalizing the multiple original attention scores to obtain attention weights at multiple time steps; matching and multiplying the bidirectional hidden state vector with the attention weights at multiple time steps to generate a weighted result at multiple time steps; and adding the weighted result at multiple time steps element-wise according to the bidirectional hidden state vector to generate the semantic feature vector.

[0076] Furthermore, the mapping module 12 is also used to perform the following steps: perform node matching based on the semantic feature vector and the manufacturing semantic knowledge graph to extract the embedding vectors of multiple nodes; calculate cosine similarity based on the embedding vectors of the multiple nodes to obtain multiple cosine similarity values; set a similarity threshold and compare the multiple cosine similarity values ​​with the similarity threshold; extract target nodes with cosine similarity greater than the similarity threshold, determine the target nodes as the first nodes that have a semantic association with the time series feature vector, and perform iterative determination based on the above process until all nodes are traversed to obtain multiple associated nodes; and match and integrate the multiple associated nodes with the multiple cosine similarity values ​​to construct the node-data association mapping matrix.

[0077] Furthermore, the optimization module 13 is also used to perform the following steps: performing manufacturing operation collaborative optimization analysis based on the multi-level manufacturing operation data model to construct manufacturing operation collaborative optimization process data; decomposing the manufacturing operation collaborative optimization process data into task scheduling optimization sub-branch, resource allocation optimization sub-branch, and exception handling optimization sub-branch, wherein the task scheduling optimization sub-branch, resource allocation optimization sub-branch, and exception handling optimization sub-branch are in parallel relationship; reading data from the multi-level manufacturing operation data model through the task scheduling optimization sub-branch to obtain production plan data and equipment availability data; performing mixed integer programming based on the production plan data and equipment availability data to generate a task priority sequence and encapsulate the equipment allocation scheme to construct task scheduling decision instructions; and using the resource allocation optimization sub-branch... The multi-level manufacturing operation data model is read to obtain material inventory data, workstation load data, and logistics path data. Based on the material inventory data, workstation load data, and logistics path data, multi-objective optimization is performed to generate material delivery paths and workstation load balancing schemes, which are then encapsulated to construct resource allocation decision instructions. The multi-level manufacturing operation data model is read through the anomaly handling optimization sub-branch to obtain equipment status anomaly data and process parameter over-limit data. Based on the equipment status anomaly data and process parameter over-limit data, hybrid reasoning is performed to generate anomaly handling schemes and early warning notification data, which are then encapsulated to construct anomaly handling decision instructions. The task scheduling decision instructions, resource allocation decision instructions, and anomaly handling decision instructions are merged and sorted to generate the intelligent decision instruction set.

[0078] Furthermore, the execution module 14 is also used to perform the following steps: sending the intelligent decision instruction set to the equipment controller and the material handling system to trigger equipment control instructions and material handling instructions; executing the equipment control instructions and the material handling instructions to collect real-time data, obtain actual equipment status change data, actual material flow data, and actual production cycle data, and using the actual equipment status change data, actual material flow data, and actual production cycle data as execution feedback data; performing execution expectation analysis based on the intelligent decision instruction set, setting expected execution targets, comparing and analyzing the execution feedback data and the intelligent decision instruction set item by item according to the expected execution targets, and calculating multiple deviation components; synthesizing the multiple deviation components according to the data direction to construct a deviation vector; calculating the length based on the deviation vector, obtaining a comprehensive length value, converting it, and generating a comprehensive deviation metric value; when the comprehensive deviation metric value exceeds a preset deviation threshold, triggering an incremental learning signal to perform a closed-loop update for manufacturing operation collaborative optimization.

[0079] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A smart manufacturing operation management method based on an industrial internet of things, characterized by, The method includes: By sensing and collecting multi-dimensional manufacturing operation data through the Industrial Internet of Things, edge processing is performed on the multi-dimensional manufacturing operation data to obtain a standardized manufacturing data stream. The standardized manufacturing data stream is mapped to a manufacturing semantic knowledge graph to construct a multi-level manufacturing operation data model. Based on the aforementioned multi-level manufacturing operation data model, collaborative optimization of manufacturing operations is carried out, and a set of intelligent decision-making instructions is formulated. The system executes the intelligent decision-making instruction set, monitors execution feedback data in real time, and performs closed-loop updates based on the execution feedback data. 2.The industrial Internet of Things based intelligent manufacturing operation management method of claim 1, wherein, Edge processing is performed on the multidimensional manufacturing operation data to obtain a standardized manufacturing data stream. The method includes: The edge gateway node of the industrial Internet of Things is located, and outlier detection is performed on multi-dimensional manufacturing operation data based on the edge gateway node to identify abnormal outliers. Based on the aforementioned outliers, the multidimensional manufacturing operation data is cleaned to obtain a cleaned dataset. The cleaned dataset is imputed by time-series interpolation to obtain a completed dataset; Based on the completed dataset, normalization processing is performed, and the normalization result is converted into a format to generate time series data. The time series data is structured and encapsulated according to data type to generate the standardized manufacturing data stream. 3.The industrial Internet of Things based intelligent manufacturing operation management method of claim 1, wherein, The method involves mapping the standardized manufacturing data stream to a manufacturing semantic knowledge graph to construct a multi-layered manufacturing operation data model, including: Based on the standardized manufacturing data stream, equipment identification analysis is performed to extract multiple equipment identification tag data; Based on the standardized manufacturing data stream, perform time feature analysis to extract time series feature vectors; Construct a manufacturing semantic knowledge graph, perform semantic association calculations between the time series feature vectors and the manufacturing semantic knowledge graph, and construct a node-data association mapping matrix; The standardized manufacturing data stream is embedded into the manufacturing semantic knowledge graph according to the node-data association mapping matrix and the multiple equipment identification tag data, thereby constructing the multi-level manufacturing operation data model. 4.The industrial Internet of Things based intelligent manufacturing operation management method of claim 3, wherein, The process of constructing a semantic knowledge graph includes the following methods: Based on the multidimensional manufacturing operation data, entity extraction is performed to construct a candidate entity set; Based on the multidimensional manufacturing operation data, relationship extraction is performed to construct a candidate relationship edge set; Perform string similarity analysis on the candidate entity set and the candidate relation edge set, and calculate the first similarity. Perform semantic similarity analysis on the candidate entity set and the candidate relation edge set, and calculate the second similarity. Based on the first similarity and the second similarity, a weighted calculation is performed, and the candidate entity set and the candidate relation edge set are merged according to the weighted result to obtain the initial merged data; Based on the initial merged data, conflict analysis is performed to extract conflict relationships. The initial merged data is then updated according to the conflict relationships to determine the entity fusion set and the relationship edge fusion set. Based on the entity fusion set and the relation edge fusion set, an initial manufacturing semantic knowledge graph is constructed by associating and filling in the data. The initial manufacturing semantic knowledge graph is subjected to quality assessment. When the quality assessment meets the preset quality indicators, the manufacturing semantic knowledge graph is generated. 5.The industrial Internet of things based smart manufacturing operation management method according to claim 3, wherein, The method involves performing semantic association calculations between the time-series feature vectors and the manufacturing semantic knowledge graph to construct a node-data association mapping matrix, including: Extract multiple node embedding vectors based on the aforementioned manufacturing semantic knowledge graph; The time series feature vector is used as the input sequence, and a bidirectional long short-term memory network is used to process the input sequence: S1: The bidirectional long short-term memory network includes a forward long short-term memory layer and a backward long short-term memory layer; S2: The input sequence is processed in chronological order through the forward long short-term memory layer to output the forward hidden state sequence; S3: The input sequence is processed in reverse chronological order through the backward long short-term memory layer to output the backward hidden state sequence; The forward hidden state sequence and the backward hidden state sequence are concatenated into vectors to generate a bidirectional hidden state vector; The bidirectional hidden state vector is calculated using an attention mechanism to generate a semantic feature vector. Based on the semantic feature vector and the manufactured semantic knowledge graph, a node-data association mapping matrix is ​​constructed. 6.The industrial Internet of things based smart manufacturing operation management method according to claim 5, wherein, The semantic feature vector is generated by calculating the bidirectional hidden state vector through an attention mechanism, and the method includes: The bidirectional hidden state vector is fully connected to perform the attention mechanism to obtain multiple original attention scores. The multiple original attention scores are normalized to obtain attention weights at multiple time points; The bidirectional hidden state vector is matched and multiplied with the attention weights at the multiple time points to generate a weighted result at multiple time points; The weighted results at the multiple time points are summed element-wise according to the bidirectional hidden state vector to generate the semantic feature vector. 7.The industrial Internet of things based smart manufacturing operation management method of claim 5, wherein, The method for constructing a node-data association mapping matrix based on the semantic feature vector and the manufacturing semantic knowledge graph includes: Based on the semantic feature vector and the manufacturing semantic knowledge graph, node matching is performed to extract the embedding vectors of multiple nodes; Cosine similarity is calculated based on the embedding vectors of the multiple nodes to obtain multiple cosine similarity values; Set a similarity threshold, and compare the plurality of cosine similarity values ​​with the similarity threshold; Extract target nodes with cosine similarity greater than the aforementioned similarity threshold, and determine the target nodes as the first nodes that have a semantic association with the time series feature vector. Based on the above process, perform iterative determination until all nodes have been traversed to obtain multiple associated nodes. The node-data association mapping matrix is ​​constructed by matching and integrating the multiple associated nodes with the multiple cosine similarity values. 8.The industrial Internet of things based smart manufacturing operation management method according to claim 1, wherein, Based on the aforementioned multi-level manufacturing operation data model, collaborative optimization of manufacturing operations is performed, and an intelligent decision-making instruction set is formulated. The method includes: Based on the multi-level manufacturing operation data model, a collaborative optimization analysis of manufacturing operations is performed to construct collaborative optimization process data for manufacturing operations. The manufacturing operations collaborative optimization process data is decomposed into task scheduling optimization sub-branch, resource allocation optimization sub-branch, and anomaly handling optimization sub-branch, which are in parallel relationship; The task scheduling optimization sub-branch reads data from the multi-level manufacturing operation data model to obtain production plan data and equipment availability data. Based on the production plan data and the equipment availability data, mixed integer programming is performed to generate a task priority sequence and equipment allocation scheme, which are then encapsulated to construct task scheduling decision instructions. The resource allocation optimization sub-branch reads data from the multi-level manufacturing operation data model to obtain material inventory data, workstation load data, and logistics path data. Based on the material inventory data, the workstation load data, and the logistics path data, multi-objective optimization is performed to generate material delivery paths and workstation load balancing schemes, which are then encapsulated to construct resource allocation decision instructions. The anomaly handling optimization sub-branch reads data from the multi-level manufacturing operation data model to obtain equipment status anomaly data and process parameter over-limit data. Based on the abnormal equipment status data and the excessive process parameter data, hybrid reasoning is performed to generate an anomaly handling plan and encapsulate the early warning notification data to construct an anomaly handling decision instruction. The task scheduling decision instructions, the resource allocation decision instructions, and the anomaly handling decision instructions are merged and sorted to generate the intelligent decision instruction set. 9.The industrial Internet of things based smart manufacturing operation management method of claim 1, wherein, The method includes: Executing the intelligent decision-making instruction set, monitoring execution feedback data in real time, and performing closed-loop updates based on the execution feedback data. The intelligent decision-making instruction set is sent to the equipment controller and the material handling system to trigger equipment control instructions and material handling instructions. The system executes the equipment control commands and the material handling commands to collect data in real time, thereby obtaining actual equipment status change data, actual material flow data, and actual production cycle data. The actual equipment status change data, the actual material flow data, and the actual production cycle data are used as execution feedback data. Based on the intelligent decision instruction set, an execution expectation analysis is performed, an expected execution target is set, and the execution feedback data and the intelligent decision instruction set are compared and analyzed item by item according to the expected execution target to calculate multiple deviation components. The multiple deviation components are combined according to the data direction to construct a deviation vector; The length of the deviation vector is calculated, and the comprehensive length value is converted to generate a comprehensive deviation metric. When the comprehensive deviation metric exceeds a preset deviation threshold, an incremental learning signal is triggered to perform a closed-loop update for collaborative optimization of manufacturing operations.

10. The intelligent manufacturing operation management system based on the industrial internet of things, characterized by, The system is used to execute the intelligent manufacturing operation management method based on the Industrial Internet of Things as described in any one of claims 1-9, and the system includes: The processing module is used to collect multi-dimensional manufacturing operation data through industrial IoT sensing, perform edge processing on the multi-dimensional manufacturing operation data, and obtain a standardized manufacturing data stream. The mapping module is used to map the standardized manufacturing data stream to a manufacturing semantic knowledge graph to construct a multi-level manufacturing operation data model. The optimization module is used to perform collaborative optimization of manufacturing operations based on the multi-level manufacturing operation data model and to formulate a set of intelligent decision-making instructions. The execution module is used to execute the intelligent decision instruction set, monitor the execution feedback data in real time, and perform closed-loop updates based on the execution feedback data.

Images